Identification of Key Gmaw Fillet Weld Parameters and Interactions Using Artificial Neural Networks

Fillet welds are one of the most commonly used weld joints but one of the most difficult to weld consistently. This paper presents a technique using Artificial Neural Networks (ANN) to identify the key Gas Metal Arc Welding (GMAW) fillet weld parameters and interactions that impact on the resultant geometry, when using a metal cored wire. The input parameters to the model were current, voltage, travel speed; gun angle and travel angle and the outputs of the model were penetration and leg length. The model was in good agreement with experimental data collected and the subsequent sensitivity analysis showed that current was the most influential parameter in determining penetration and that travel speed, followed closely by current and voltage were most influential in determining the leg length. The paper also concludes that a ‘pushing’ travel angle is preferred when trying to control the resultant geometry mainly because both the resultant leg length and penetration appear to be less sensitive to changes in heat input. Introduction Presently there is no economic technology available to accurately measure the actual internal geometry of a fillet weld without destructively testing the work piece. The external geometry of a fillet weld can be measured easily using specifically designed gauges, but the internal characteristics, such as penetration cannot be measured as easily. Penetration is critical in determining the structural integrity of a fillet weld to ensure that the axis between the bar and the plate is effectively ‘cut’. In order to guarantee satisfactory penetration and weld geometry it is imperative that a high level of control of the welding parameters can be demonstrated. Over the years there have been numerous studies [1, 3 & 4] proving that the ability to predict weld geometry is related to the level of control of the parameters. Miller [1] reported that tight control of electrode placement; fit-up, welding position and welding procedures are required to ensure repeatability. Initial investigations would seem to indicate that increasing stick out increases spatter but reduces penetration and the width of the weld bead. There are also studies [12,22] which demonstrate that alternating the shielding gas can have a positive effect on the level of weld penetration whilst reducing leg length and also that the shielding gas flow rate can be reduced substantially without impacting the overall coverage and quality of the weld. Tham et al [3] also demonstrated the correlation between the welding parameters and the resultant bead geometry. Welder Current (A) Volts (V) Heat Input (kJ/mm) Assumed average travel speed 400mm/min 1 204 20.8 0.636 5 224 22.1 0.743 6 238 19.8 0.707 7 236 22 0.779 8 212 21.5 0.684 10 234 22.9 0.804 12 240 24.8 0.893 15 229 24.4 0.838 18 224 22.8 0.766 19 215 24.6 0.793 Average 225.6 22.57 0.764 Min 204 19.8 0.636 Max 240 24.8 0.893 Variation (%) 15.0% 20.2% 28.7% Table 1: Variation in parameter settings for manual welding Table 1 shows the results of a short study of a number of welders showing the parameters they used to complete a series of downhand fillet welds. The variation seen in this study highlights that even within a group of experienced welders there is a high level of variation of the input parameter settings for a relatively simple fillet weld arrangement. There have been numerous papers written and studies undertaken on the subject of controlling GMAW weld parameters and resultant geometry however as figure 1 shows, the large number of input parameters and variables (this list is indicative not exhaustive) makes it extremely challenging to understand exactly what impact the variation each of the inputs (and their interactions with each other) has on the resultant fillet weld. The impact of this variation will be discussed later. However, in order to maintain consistent quality fillet welds it is critical to understand the extent to which each of these input parameters (and their interactions) affect the resultant outputs. Furthermore if a robust process control model can be developed that can demonstrate tight control of the parameters and interactions that affect the joint geometry, then confidence can be increased that sufficient penetration and leg length is being achieved whilst heat input and distortion is minimised. This paper details the 1st stage of a wider scope of work which will focus on understanding how the input parameters in figure 1 interact and impact the resultant fillet weld geometry. One of the key goals of this research is ultimately to provide guidance on parameter control to ensure that all automated welding is carried out consistently. This paper however will deal specifically with understanding the impact and interactions the following parameters: current voltage, travel speed, travel angle and gun angle, have on the resultant fillet weld geometry (leg length and penetration). There are many sources of guidance on input parameter selection for GMAW, in both academic and industrial publications. However on closer inspection the wealth of guidance on offer can be confusing and at times contradictory. The following examples, taken from a mixture of supplier’s websites, technical documentation and academic publications, highlight the level of variation and the complexities involved in trying to identify exactly what the optimum gun and travel angles are for GMAW fillet welding. Miller Electric [6] advise that a ‘pushing’ (+ve) travel angle produces less penetration and a flatter bead (so conversely a ‘pulling’ (-ve) travel angle produces a deeper/narrower bead). Miller Electric [6] also advises using a travel angle of 5°-15° because increasing to greater than 20°-25° creates more spatter, less penetration and is consequently less stable. Similar advice can be found from Esab’s online handbook where a backhand (pulling) technique is recommended to reduce spatter and produce a more stable arc. Esab also advise that a backhand technique increases penetration and bead width, whereas a forehand (pushing) technique reduces the penetration and bead width of the resultant weld. BOC [7] advises that for metal cored GMAW the travel angle should be 20°-30° (pushing). Harwig [8] advises that higher deposition rates can be achieved with a 15° ‘pushing’ travel angle, Bhattacharya [9] advises that in general ‘pushing’ reduces deposition efficiency, however Lincoln Electric [10] advise using a ‘pulling’ angle of between 20°-30°. The range of gun angles also varies depending on what publication is being referred to. Lincoln Electric [10] recommends using a gun angle of less than 45° and BOC [7] a gun angle range of 30°-40°. Tham et al [3] also conducted investigations using a fixed gun angle of 45°. The experiments detailed within this paper, with the aid of an ANN model aim to try and provide some clarity as to what guidance can be confidently applied to GMAW mild steel fillet joints (6mm). Artificial Neural Networks (ANNs) are computing systems consisting of a collection of interconnected processing elements which are able to represent complex interactions between process inputs and outputs, such as that shown for fillet welding. During the model development a number of different network topologies were assessed including Multilayer Perceptron (MLP), Generalised Feed Forward (GFF) and Probabilistic Neural Network (PNN). As part of the model development the software produced a report comparing the accuracy of the different various network topologies. This report concluded that a Multi-Layer Perceptron (MLP) Model, with 5 inputs, 2 hidden layers and 3 output layers was the most accurate model and so was selected. ANN’s can be used to predict the outputs to a process as long as sufficient data is created and fed in to train the model. The ANN can identify patterns, trends and interactions that are too complex to be detected by other existing methods and technologies. Bhadeshia [19] suggests that ANN’s are ideal for determining welding process parameters such as penetration. ANN’s which could accurately predict the penetration and internal geometry of a fillet joint would provide a great benefit by greatly reducing the cost (material and labour) or trialling and testing new welding procedures and processes. The main benefits that ANN’s provide are:  They do not require any predefined relationship between the variables to be understood  They allow patterns, trends and interactions to be identified that otherwise would be impossible to detect.  They work well when there are a large number of diverse variables to analyse.  They can be used and applied to a variety of problems (not specific to thermo-mechanical engineering related processes)  They can be used to process symbolic data as well as numeric data. There are however some important limitations in using MLP ANN models that need to be understood.  They do not explain why patterns and/or interactions exist so it requires analyses and interpretation of the results  They may not always find the optimal solution  The model development requires an element of trial and error (trying different network topologies, iterations, number of layers...etc.) in order to try and create the most accurate model. There are numerous examples of ANN’s that have been developed to predict GMAW fillet weld geometries. [11-18] provide examples of ANN’s that have been successfully developed using a subset of the input and output parameters shown in Figure 1. However there are no publications that investigate the impact of both travel angle and the gun angle (and their interactions) have on the resultant fillet weld geometry (horizontal leg length, vertical leg length and penetration). This paper will use ANNs to analyse the relationship/impact that the current, voltage, travel speed, torch travel angle and gun angle have on the resultant fillet weld geometry (leg length and penetration). It will also analyse if the interactions between these input parameters are significant in influencing the resultant weld geometry. Figure 1: Fillet Weld Inputs and Outputs Experimental Procedure In total 97 test plates were welded on the rig (figure 2) at Strathclyde University using a customised jig to set the gun and travel angle. The jig was designed to allow the torch (gun) angle (figure 3) to be set at 5° increments from 35° 50° relative to the horizontal base plate. The jig also allowed the torch travel angle (figure 4) to be set a 15° increments from -30°to +30° relative to the direction of travel. Figure 2: Image of Welding Rig Figure 3: Diagram showing gun angle orientation Direction of travel